Stay connected with KayaToday, follow us on Instagram and Facebook for the latest news and reviews delivered straight to you.
The dominant mental model for AI agents right now is one person, one assistant. You prompt it, it responds, you iterate. That framing is already becoming obsolete. At VB Transform 2026, James Zou, associate professor of biomedical data science at Stanford University, presented something far more disorienting: a system of 37,000 AI agents structured like a pharmaceutical company, which autonomously designed a cancer drug therapy that Merck later independently developed and validated, earning FDA breakthrough designation in the process.
This is not a chatbot story. It is a story about what happens when AI stops being a tool you use and starts being an organisation you manage.
From Five Agents to a Full Corporate Structure
Zou’s work began modestly. His team built what they called a Virtual Lab, a cluster of five to eight agents mirroring the structure of his physical Stanford research group. One agent played the role of principal investigator. Others acted as specialist students. They held regular group meetings and were even enrolled in what Zou described as an agent school, where supervised fine-tuning sharpened their domain expertise.
The early results were already striking. The virtual lab designed new nanobody proteins targeting recent COVID-19 variants. When those designs were tested in wet-lab conditions, they outperformed previously human-designed nanobodies in binding to the virus. That wet-lab validation gave the team confidence to scale dramatically.
The expanded system, which Zou calls the Virtual Biotech, replicates the divisional structure of a large pharmaceutical company. A Chief Scientific Officer agent oversees distinct divisions covering target discovery, molecule design, and clinical trial analysis. Within each division, agents specialise further. In the target discovery division alone, one agent focuses on genetics data, another on genomics and single-cell data, and so on down the hierarchy. The total headcount, in agent terms, reaches tens of thousands.
Why Thousands of Specialists Beat One Omniscient Model
The obvious question is whether this complexity is necessary. As foundation models grow more powerful, why not simply throw more compute at a single model rather than coordinating thousands of smaller ones?
Zou’s team ran a direct comparison. A multi-agent team and a single agent were given the same scientific problem. The multi-agent system consistently produced better, more resilient solutions. The reason, Zou argued, is that disagreement itself is productive. “In these scientific virtual labs, the agents actually get into debates and disagreements. They have to convince the other AI scientists of their ideas, and all of that elicits much more creative and robust reasoning compared to if you have a single model trying to do the problem by itself from scratch,” he said.
This maps onto something intuitive from human research culture. Peer review, lab meetings, and cross-disciplinary collaboration exist precisely because individual experts have blind spots. Replicating that friction artificially, through agent debate, appears to generate similar benefits in error-correction and creative output.
The Infrastructure Problem Nobody Talks About
Scaling to tens of thousands of agents creates an orchestration challenge that is easy to underestimate. The bottleneck is not model capability. It is data access. Legacy scientific databases were built for human researchers or pre-AI algorithms, and they do not translate cleanly into agent-readable formats. Simply wrapping an existing database with a Model Context Protocol layer, a common enterprise shortcut, does not solve the underlying mismatch.
“Even if you wrap an MCP around the existing databases and APIs, that doesn’t solve the underlying problem: the interface and APIs are not suitable for agents,” Zou said. Dropping a PDF of a research paper into an agent’s context window, for instance, is inefficient, and standard text models struggle to parse complex figures and tables accurately, which increases the risk of hallucinations compounding across thousands of agent interactions.
To address this, Zou’s team built a platform called Paperclip. Rather than forcing agents to navigate brittle, database-specific APIs, Paperclip digitises unstructured scientific data and maps it into a unified, AI-native virtual file system. Agents can then access knowledge from millions of papers using standard file-system operations, the kind of task modern large language models handle well because of their code-writing and file-navigation capabilities. Zou reported that Paperclip reduced both time and cost by more than an order of magnitude compared to running agents without this AI-native infrastructure, while also improving accuracy.
For enterprise technology teams in Singapore and Malaysia exploring agentic AI deployments, this is arguably the most practically transferable insight from Zou’s work. The data layer is not a secondary concern to be solved after the agents are built. It is the primary constraint that determines whether a multi-agent system functions at all.
The Merck Confirmation and What It Actually Proves
The most arresting result came from the clinical trial analysis component. The Virtual Biotech deployed 37,000 clinical trial agents to synthesise fragmented trial data from across the literature. Those agents identified single-cell features that predict trial success, and drugs supported by those features were approximately 50% more likely to reach market than comparable drugs without them.
Acting on this analysis, the system autonomously designed an antibody-drug conjugate targeting the CD276 protein for lung cancer. Critically, the agents completed this design using only data published before January 2025, with no access to subsequent research.
Several months later, Merck independently developed and validated the same therapeutic design. That design received FDA breakthrough designation. Zou described this as “a third-party external validation of the therapeutic design provided by the virtual biotech agents,” and the framing is careful but significant. Merck was not collaborating with Stanford. They arrived at the same answer independently, which is the closest thing to a controlled experiment you can run on a system like this.
It does not prove the agents are infallible, and Zou did not claim that. But it does suggest the system is generating scientifically coherent outputs that hold up under real-world scrutiny, which is a meaningfully higher bar than benchmark performance on curated test sets.
Managing Environments, Not Workflows
Zou closed with a conceptual shift that matters as much as the technical architecture. As multi-agent systems scale, the instinct is to design tighter workflows, to specify exactly what each agent should do and when. Zou argued this is the wrong frame entirely.
The distinction he drew is between workflows and environments. Workflows tell agents what to do. Environments provide infrastructure, guardrails, and incentives, then leave the problem open. “In workflows, we’re trying to tell agents what to do and how to do their job. But in environments, we’re providing the infrastructures, the incentives, and the guardrails, but otherwise we leave it open to incentivize agents to collaborate,” he said.
At the individual agent level, improvement comes through reinforcement learning or fine-tuning. At the system level, the lever is the environment itself. “We’re not actually fine-tuning and changing the individual models anymore, but we’re optimizing the environment,” Zou explained. “The environment itself is the object that we optimize to improve the agents.”
That reframing has consequences well beyond drug discovery. Any organisation considering large-scale agentic deployment, whether in financial services, legal research, or manufacturing, will eventually face the same design question. The answer Stanford is arriving at is that the job of human leaders in these systems is less about directing individual agents and more about architecting the conditions under which thousands of them can productively disagree, specialise, and converge on solutions. That is a genuinely new kind of management problem, and the Merck validation suggests it is one worth taking seriously.
